AI Agents Are Failing and It's Almost Never the Model's Fault | Alberto Pan, Denodo

What It Covers

Alberto Pan, CTO of Denodo, presents findings from an 850+ data leader survey revealing that enterprise AI agent failures stem primarily from data architecture problems — not model limitations. Pan explains how a universal semantic layer addresses real-time data access, semantic consistency, and governance enforcement across distributed enterprise data sources.

Key Questions Answered

• •AI Failure Root Cause: Nearly 70% of surveyed organizations report that lack of real-time data is the primary cause of AI agent failures — not model intelligence gaps. Organizations already in production report this problem at even higher rates, meaning the issue compounds as deployments scale beyond pilots into live business workflows.
• •Data Source Scale: Enterprise AI agents require data from an average of 400 sources, with over 85% of organizations needing more than 100 sources. Architects should plan data access infrastructure assuming agents will need broad, unpredictable data access — not narrow, predefined inputs — because agents by definition handle scenarios no fixed workflow can anticipate.
• •Semantic Layer Requirement: Without a defined semantic layer, even the most capable AI model is forced to guess business context — producing hallucinations. Organizations should explicitly encode business rules, KPI formulas, and terminology mappings (e.g., what constitutes a hospital readmission) into a centralized semantic layer before deploying agents into production workflows.
• •Avoid the Ad Hoc Trap: Building isolated data layers per AI application creates siloed agents that cannot collaborate in multi-agent architectures. Pan recommends establishing one shared data infrastructure with consistent governance, semantics, and security policies across all agents — rather than delivering quick wins through per-application data shortcuts that become dead ends at scale.
• •Adoption Timeline: Pan estimates that by end of 2027, roughly 25% of large enterprises will run agentic AI across approximately 20% of critical production workflows. Data maturity is the differentiating factor — organizations with established data product foundations already show significantly higher AI production adoption rates than peers with fragmented architectures.